Deep Learning for Hail Size Estimation Using Polarimetric Radar Data

In 2022, convective storms producing large hailstones were responsible of significant damages in Europe, particularly in France, where the insurance sector was severely hit. As a result, prices for 2023 might rise significantly compared to 2022, leaving a lot of pressure on the sector. Thus, the need to correctly assess and measure hail damage in real time has become crucial to assist communities in Europe and help them recover from such events.

The potential of radar data, when it comes to the detection of hail, is not to be challenged anymore: radar remains since decades the best remote sensing tool to detect such hazard. As a result, a lot of computational methods using radar data were developed to detect hail and estimate its size in real-time. Probability of hail based algorithms, hydrometeor classification using fuzzy logic or new methods using machine learning were explored to locally quantify the presence of hail in storms. Recent studies even suggest that studying the structure of the parent storm when there is hail could be of interest. Deep Learning algorithms built to analyze images and detect features in multi-dimensional data appear to be a good lead for estimating the occurrence and the size of hail in storms and on the ground. Applied to radar data, such techniques could detect signatures specific to hail and enhance our understanding on why one storm can produce 2cm hailstones and another 10cm hailstones.

This work aims to present a novel approach using Convolutional Neural Networks (CNNs) on radar images to detect hail occurrence and estimate its size. Based on multiple hail cases in France from 2017 to 2022 and hail reports on the ground, a CNN is trained and tested. Comparisons are made with existing sizing algorithms such as Fuzzy-Logic Hydrometeor Classification or Maximum Estimated Size of Hail (MESH). First results show great perspectives, even with really simple network architectures.